Classification of suspicious lesions on prostate multiparametric MRI using machine learning

Area of Science:

• Radiology
• Medical Imaging
• Oncology

Background:

• Accurate classification of clinically significant prostate cancer is crucial for effective treatment.
• Multiparametric magnetic resonance imaging (mpMRI) offers detailed insights into prostate tissue characteristics.
• The PROSTATEx challenge aimed to advance automated prostate cancer detection and classification.

Purpose of the Study:

• To develop and evaluate a radiomics-based approach for classifying clinically significant prostate cancer.
• To compare the performance of different machine learning algorithms in this classification task.
• To assess the effectiveness of radiomic features extracted from mpMRI for prostate cancer detection.

Main Methods:

• A radiomics workflow was developed using the MIM imaging platform for image analysis.
• Manual segmentation of prostate and peripheral zones was performed on T2-weighted images.
• Rigid image fusion was applied to T2-weighted, proton density-weighted, dynamic contrast-enhanced, and diffusion-weighted MRI sequences.
• Radiomic features (intensity and texture) were extracted and characterized using nine histogram descriptors.
• Classification was performed using Classification and Regression Trees (CART), Random Forests, and Adaptive LASSO.

Main Results:

• The Random Forests model achieved the highest Area Under the Curve (AUC) of 0.82 on the test dataset.
• This AUC of 0.82 ranked fourth among 71 entries and was the top score for feature-based methods.
• CART and Adaptive LASSO achieved AUCs of 0.76.
• A total of 216 radiomic features were extracted and analyzed.

Conclusions:

• Radiomics analysis of multiparametric MRI holds significant potential for classifying clinically significant prostate cancer.
• The Random Forests algorithm demonstrated superior performance in this classification task.
• The developed radiomics approach achieved competitive results in the PROSTATEx challenge, highlighting its clinical relevance.